Adaptive stride prefetching for the secondary data cache of UMA and NUMA

Prefetching is a promising approach to tackle the memory latency problem. T wo basic variants of hardware data prefetching methods are sequential prefe tching and stride prefetching. The latter based on stride calculation of fu ture references has the potential to out-perform the former which is based on the data locality. In this paper, a typical stride prefetching and its i mproved version, adaptive stride prefetching, are compared in quantitative way using simulation for some parallel benchmark programs in the context of uniform memory access and non-uniform memory access architectures. The sim ulation results show that adaptability of stride is essential since the pro posed adaptive scheme can reduce pending stall time which is large in the t ypical scheme.